Card: Krill naming/install fix + shared benchmark comparison table (partial)

README.md

@@ -1,80 +1,79 @@
 ---
 license: gemma
 library_name: mlx
+base_model: yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF
+base_model_relation: quantized
+pipeline_tag: text-generation
+model_type: gemma4_unified
 tags:
   - nvfp4
   - mlx
   - krill
   - gemma4
-  - code
+  - gemma4_unified
+  - apple-silicon
   - agentic
-base_model: yuxinlu1/gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2-GGUF
-pipeline_tag: text-generation
+  - tool-use
 ---
 
 # gemma-4-12B-agentic-fable5-composer2.5-v2 — NVFP4 (MLX)
 
 > Original fine-tune by [yuxinlu1](https://huggingface.co/yuxinlu1) (`gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2`); this repo is an NVFP4 (MLX) requant for fast local inference on Apple Silicon.
 
-A mixed-precision **NVFP4** conversion of the Gemma‑4‑12B *agentic* fine‑tune. The weights are
-plain MLX safetensors; it was built for and is tested with **Krill** (a pure Swift + MLX
-inference engine for Apple Silicon). Loads in **~1.7 s** at **6.8 GB**.
+A mixed-precision **NVFP4** conversion of the Gemma‑4‑12B *agentic* fine‑tune (optimized for tool‑use / τ²‑bench). The weights are plain MLX safetensors;
+built for and tested with **[Krill](https://github.com/srvsngh99/Krill)** — a pure Swift + MLX
+inference engine for Apple Silicon (no Python, no GGUF at inference). Loads in **~1.7 s** at **~6.8 GB**.
 
 ## The format
 
-- The bulk of the weights are **NVFP4** (4‑bit float, group_size 16); the attention `o_proj`
-  and the vision/audio projectors are kept at **8‑bit affine** (group_size 64). This "protected"
-  mixed recipe recovers the quality uniform 4‑bit loses on those sensitive modules while keeping
-  4‑bit speed and size. Same recipe as
-  [`srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4`](https://huggingface.co/srv-sngh/gemma-4-12B-coder-fable5-composer2.5-nvfp4).
-- **Conversion path:** plain‑bf16 safetensors → key‑remap to MLX layout → NVFP4 requant. No GGUF
-  round‑trip (so none of GGUF's k‑quant quality loss).
+- Bulk weights are **NVFP4** (4‑bit float, group_size 16); attention `o_proj` and the vision/audio
+  projectors are kept at **8‑bit affine** (group_size 64). This "protected" mixed recipe recovers the
+  quality uniform 4‑bit loses on those sensitive modules while keeping 4‑bit speed and size.
+- **Conversion:** bf16 safetensors → key‑remap to MLX layout → NVFP4 requant. No GGUF round‑trip.
 
 ## Compatibility (please read)
 
-This is an **MLX** checkpoint, not GGUF and not a HF/transformers checkpoint. To run it an engine
-needs (1) the `gemma4_unified` architecture (text+vision+audio) and (2) support for the
-mixed‑precision NVFP4 config (top‑level nvfp4 block + per‑module 8‑bit overrides keyed by path).
-
-- ✅ **Krill** — runs natively today (the recipe was designed for its loader).
-- ⚙️ **Other MLX engines** — possible if they add `gemma4_unified` + this quant schema; not
-  drop‑in with vanilla `mlx_lm`/`mlx_vlm` as‑is.
-- ❌ **llama.cpp / Ollama** — no (GGUF only).
-- ❌ **transformers / vLLM** — no (these want bf16 or their own quant formats).
+This is an **MLX** checkpoint — not GGUF, not a HF/transformers checkpoint. To load it an engine needs
+(1) the `gemma4_unified` architecture (text+vision+audio) and (2) the mixed‑precision NVFP4 config
+(top‑level nvfp4 + per‑module 8‑bit overrides). Today that means **Krill**; it is **not** drop‑in for
+vanilla `mlx_lm`/`mlx_vlm`, and **not** loadable by llama.cpp/Ollama (GGUF) or transformers/vLLM.
 
-## Run it with Krill
+## Install Krill & run
 
 ```bash
-krill pull srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4
+# Homebrew:
+brew tap srvsngh99/krill && brew install krill
+# …or one-line installer (Apple Silicon):
+curl -fsSL https://raw.githubusercontent.com/srvsngh99/Krill/main/install.sh | sh
+
+krill pull srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4   # by full path (alias TBD)
 krill run  srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4 "Write a Python LRU cache."
-# OpenAI-compatible server:
-krill serve --model srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4 --port 57455
-# enable the reasoning channel:
-KRILL_ENABLE_THINKING=1 krill run srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4 "..."
+krill serve --model srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4 --port 57455     # OpenAI-compatible API
+KRILL_ENABLE_THINKING=1 krill run srv-sngh/gemma-4-12B-agentic-fable5-composer2.5-v2-nvfp4 "..."   # reasoning channel
 ```
 
 ## Benchmarks
 
-All numbers are **pass@1 / accuracy**, single greedy pass, run with Krill on an M4 Pro. All three
-models are in this **same NVFP4 format** (true apples‑to‑apples). HumanEval+/MBPP+ are
-[EvalPlus](https://github.com/evalplus/evalplus) (stricter tests); MBPP uses the 378‑problem
-EvalPlus set; GSM8K is a 150‑problem subset, 8‑shot. **Not** EvalPlus‑leaderboard‑comparable.
+**pass@1 / accuracy**, single greedy pass, run with Krill on an M4 Pro. All three models are in this
+**same NVFP4 format** (true apples‑to‑apples). HumanEval+/MBPP+ are
+[EvalPlus](https://github.com/evalplus/evalplus) (stricter tests); MBPP = the 378‑problem EvalPlus set;
+GSM8K = 150‑problem subset, 8‑shot. **Not** EvalPlus‑leaderboard‑comparable.
 
 | Model | mode | HumanEval | HumanEval+ | MBPP | MBPP+ | GSM8K |
 |---|---|---|---|---|---|---|
 | Google gemma-4-12B-it (base) | off | 57.3 | 56.7 | 42.1 | 37.6 | 95.3 |
-| Google gemma-4-12B-it (base) | on | 48.8 | 48.8 | 49.5 | 43.9 | — |
-| agentic v2 (this model) | off | 83.5 | 81.7 | — | — | — |
-| agentic v2 (this model) | on | 86.0 | 82.9 | — | — | — |
+| Google gemma-4-12B-it (base) | on | 48.8 | 48.8 | 49.5 | 43.9 | 90.7 |
+| coder v1 | off | 81.7 | — | 79.4 | — | 90.7 |
+| **agentic v2** ⟵ this model | off | 83.5 | 81.7 | — | — | — |
+| **agentic v2** ⟵ this model | on | 86.0 | 82.9 | — | — | — |
 
-> ⚠️ **Partial results — this run is still in progress.** Empty cells (—) are filling in; the card will be updated as the full sweep (coder + remaining agentic/GSM8K) completes.
+> ⚠️ **Partial — benchmark run still in progress.** Empty cells (—) are filling in; this card updates as the full sweep completes.
 
-**Read so far:** the code/agentic fine‑tunes massively out‑code the Google base (HumanEval ~83% vs
-57%), while the base is stronger at math (GSM8K). Reasoning‑on helps the fine‑tunes but tends to
-*hurt* the base model's coding (it over‑reasons and mangles the code block). Decode ≈ **28 tok/s**.
+**Takeaways:** the code/agentic fine‑tunes massively out‑code the Google base on HumanEval/MBPP, while
+the base is stronger at math (GSM8K). Reasoning‑on helps the fine‑tunes but tends to *hurt* the base's
+coding (it over‑reasons and mangles the code block). Decode ≈ **28 tok/s**.
 
 ## Credits & license
 
-Fine‑tune © its original author ([yuxinlu1](https://huggingface.co/yuxinlu1)); base model is
-Google Gemma 4, under the [Gemma license](https://ai.google.dev/gemma/terms). This repo only
-changes quantization/packaging.
+Fine‑tune © its original author ([yuxinlu1](https://huggingface.co/yuxinlu1) (`gemma-4-12B-agentic-fable5-composer2.5-v2-3.5x-tau2`)); base model is Google Gemma 4, under the
+[Gemma license](https://ai.google.dev/gemma/terms). This repo only changes quantization/packaging.